Deep Learning For Face Alignment

Face alignment task is one important component in the face recognition system. It's the basis for many face related problem, such as 3D face recon-struction, face identification. Common methods are based on cascated regres-sion, which builds regressors on hand-crafted features and refine shapes gradu-ally. Although the performance of face alignment problem has been improved greatly and even applied on actual products, face alignment under unconstrained circumstances still face many challenges, such as pose, expression, and occu-lusion. Recently, deep learning models have been applied to solve many visual tasks and achivevd state-of-the-art performance.In this paper, we develop a new face alignment method based on deep learning models. We solve the face alignment problem by two stages. The first stage is response mapping stage, for each facial key point, we learn one response map by deep learning model. We enhance the fully convolutional network to suit the face alignment problem. Firstly, we define the response maps and KL divergence loss function. Secondly, we replace the common used max pooling with convolutional pooling.The second stage is shape mapping stage, which deals with response maps to get final face shape. The shape generated from enhanced fully convolutional network (EFCN) contains much noise because the EFCN doesn't model the relation between points explicitly. We explored two ways of reducing noise,PCA and deep autoencoder and find that deep autoencoder is better than PCA.We have done multiple experiments on 300W dataset to evaluate the per-formance of our methods, and achieve mean error rate of 4.32%, which outper-form the state-of-the-art method by a large margin (CFSS, 5.76%).